Dario witnessed the divergence of Baidu's AI路线.
By Cheng Yanjing Edited by Hu Min
Anthropic founder Dario's career start was somewhat dramatic. In 2014, just out of school, he joined Baidu and found himself in the middle of the first "AI academic faction vs. AI engineering faction" battle in Chinese internet history.
At the time, the Silicon Valley AI academic star team, led by Andrew Ng, Han Xu, Dario, Wu Shuang, and others, went head-to-head with the Xierqi search AI engineering army led by Wang Haifeng and Jia Lei. The two sides conducted an internal test on the speech recognition track, and Jia Lei's team won by a narrow margin.
But the final outcome took a shocking turn: the better-performing Jia Lei was forced out. Andrew Ng turned the tables and took over Jia Lei's speech team.
What exactly did Dario experience in Baidu's AI system that year? Why did Baidu develop two AI routes? Why did the outcomes of Andrew Ng and Jia Lei reverse? (This article's author has long focused on AI stories at Baidu and other internet giants. Feel free to add WeChat Who123start for more insider details.)
01 Baidu AI Split in Two: Wang Haifeng and Yu Kai "Part Ways"
To understand this clash, we must go back to a major organizational restructuring at Baidu.
In 2013, all of Baidu's AI resources were concentrated in the group's basic technology system. This team had two aces: Wang Haifeng, who later became Baidu's CTO, and Yu Kai, now founder of Horizon Robotics. The two had known each other for years and together built Baidu's IDL (Deep Learning Lab). In terms of business division, Wang Haifeng was responsible for NLP and the speech lab under the basic technology system, while Yu Kai focused on computer vision and multimedia.
The turning point came in 2014 with an organizational restructuring: Wang Haifeng and his team were moved from the basic technology system into the "Search Business Group," and his reporting line changed from Wang Jin to Xiang Hailong. Yu Kai remained in the independent AI system, and IDL continued to focus on papers, vision models, and cutting-edge research, more akin to a research institute.
From then on, Baidu AI was split in two: one side stayed in the group technology system, the other entered the search business department. Former comrades in the same trench thus diverged, becoming rivals to compete with in the future.
At that time, Wang Haifeng's team was star-studded, forming the basic lineup of later Baidu AI technology system: including Wu Hua (now Chairman of Baidu Technology Committee), Jia Lei (now Head of Baidu Application Model R&D Department), Wu Tian (Vice President of Baidu Research Institute), Zhao Shiqi (former GM of Baidu Search), Yu Dianhai (now Chief Architect of Baidu Large Model Infrastructure), Wu Haifeng (now Head of Douyin Search), Dai Shuaixiang (founder of Moran Cognition), etc.
Meanwhile, Yu Kai was making high-profile moves, secretly conducting his plan to recruit Geoffrey Hinton, intending to bring in the father of deep learning.
02 Baidu's Parallel Universe: Dario, Hinton, and Ilya Almost Worked Together
Yu Kai and Geoffrey Hinton were old acquaintances, having cooperated once at ICML in Montreal. After Hinton won the ImageNet championship in 2012 with AlexNet, Yu Kai immediately wrote Hinton an email expressing interest in collaboration.
Then, after discussions among Robin Li, Wang Jin, and Yu Kai, Baidu was willing to offer $12 million to buy the company co-founded by Hinton and his students Ilya Sutskever and Alex Krizhevsky.
However, this acquisition quickly turned into a bidding war. Baidu's Yu Kai, Google's Jeff Dean, Microsoft's Deng Li, and Deepmind's Hassabis all participated in the bidding, where identities were hidden. Deepmind and Microsoft withdrew successively, leaving only Baidu and Google bidding, driving the price up to $44 million. Ultimately, Hinton cited a lumbar injury making long trips to China inconvenient and chose to join Google.
If the acquisition had succeeded, Baidu would have brought in both OpenAI co-founder Ilya and Anthropic founder Dario around the same time.
Although the Hinton team acquisition fell through, it showed the American academia and industry Baidu's vision and determination in AI.
In April 2014, at a Starbucks on Castro Street in Mountain View, Silicon Valley, Yu Kai, who already had entrepreneurial ideas, tried to persuade Stanford professor Andrew Ng to join Baidu and lead Baidu AI to new heights. After half a year of persistent persuasion, Andrew Ng officially joined.
Meanwhile, Yu Kai's journey at Baidu came to an end. At the quarterly director's meeting, Yu Kai formally resigned. After the handover, Andrew Ng gradually took over the group's AI work.
Thus began the confrontation between Baidu's AI academic faction and AI engineering faction.
03 Baidu Fires the First Shot: "AI Academic Faction vs. AI Engineering Faction"
Around 2014, at Baidu's Silicon Valley lab in Sunnyvale, California, Andrew Ng launched an aggressive hiring strategy: high compensation, high titles. His student Adam, fresh out of Stanford with a PhD, was placed at a director-level position, reportedly with an annual salary of $2 million. Dario was among them.
By the end of 2014, the Silicon Valley lab had developed an advanced speech recognition system, Deep Speech, led by Andrew Ng, current WeRide CEO Han Xu, Anthropic founder Dario, and Nvidia autonomous driving chief engineer Wu Shuang.
Unlike mainstream speech recognition solutions at the time, Deep Speech used an end-to-end deep learning architecture and outperformed Google, Bing, and Apple APIs on public datasets. However, behind the apparent glamour, the Silicon Valley team had a clear weakness: lack of real-world data, especially high-quality search data. Results remained mostly at the lab validation stage; though the papers were impressive, they were far from large-scale product deployment.
A researcher who worked in AI departments at multiple big companies told Leiphone: "At big companies like BAT, it's common for business units not to open large-scale data to AI labs."
In contrast, Jia Lei's speech recognition team under Baidu's Search Business Group faced a completely different situation. Because the search AI team was closely tied to Baidu's core businesses, they held the most real user speech and knowledge graph data. Their algorithms were simple but engineering solid, focusing on stability, performance, and engineering efficiency.
In those years, Jia Lei was one of Baidu's most prominent technical leaders. An industry saying went: if Jia Lei is not the number one in domestic speech engineering, no one dares to claim the title.
So Baidu faced a choice: the Silicon Valley lab's cutting-edge speech algorithms or the local search department's mature speech system. In the initial confrontation, the engineering faction's performance temporarily took the lead.
But according to insider sources, Andrew Ng complained to senior management that the engineering faction's data had problems, tilting the already leaning scales toward a different outcome.
Ultimately, the speech recognition team under the Search Business Group was unified under Andrew Ng's reporting. People on this line began to diverge: Jia Lei left and co-founded OrionStar with Fu Sheng. The "greenhorn" Dario, who had witnessed the bloody storm, joined Google.
Thus, Baidu's AI system, split in two, reunited on speech recognition.
According to those involved, the competition between the local and Silicon Valley factions was fierce, with colleagues working around the clock and even working while ill. The Silicon Valley lab especially needed to justify its sky-high salaries.
Although they were competitors, there was mutual respect. A core member of Jia Lei's team later recalled to Leiphone: from the perspective of today's large models, Andrew Ng was already using end-to-end speech structures back then, which was very forward-looking. Meanwhile, members of Andrew Ng's team believed that the local faction's data cleaning and processing capabilities were on par with Google and Microsoft.
After Andrew Ng absorbed the team, Deep Speech did not massively replace the original speech system within Baidu. Baidu's online products continued to evolve mainly along the original engineering system.
Meanwhile, Andrew Ng's controversial management style began to show. In 2015, Baidu publicly apologized for the ImageNet submission violation. Subsequently, the investigation results published under Andrew Ng's name confirmed that Wu Ren was fired. This decision sparked considerable controversy in the industry. Several scholars, including Tang Xiao'ou, supported Wu Ren, arguing that the violation occurred during the "informal verification period" after the competition deadline, and the results were not included in the official ranking—it was technical verification, not competition fraud, and did not warrant firing.
After this incident, some in the industry labeled Andrew Ng as "exploiting the situation to purge dissidents." Additionally, despite internal opposition, he recruited Lin Yuanqing from NEC Lab to take over as director of Baidu IDL.
These practices led to significant controversy for him at Baidu. Eventually, in March 2017, Andrew Ng left Baidu. A few months later, Baidu's Silicon Valley lab was integrated into autonomous driving and other departments.
Subsequently, Wang Haifeng took over Andrew Ng's AI team, integrating the big search AI, research institutes, and IDL into one system, establishing the AIG business unit with Wang Haifeng as head.
Four years later, Baidu AI formally merged from two separate paths into one.
04 Baidu AI: The Only Constant is Change
After the system was reconverged, personnel began to return. Baidu re-hired Jia Lei, and some former team members gradually returned. However, as AIG (Technology Platform System) gradually separated from core businesses like big search and later MEG, Wang Haifeng's team's advantage in rapid iteration based on real-world scenarios also faced new challenges.
Therefore, Baidu experienced the integration of AIG and ACG (Baidu Cloud), and then a split...
Within large companies, AI systems have always gone through cycles of division and integration, and this is not uncommon.
Now, the past tensions have subsided. The route clash that Dario experienced remains a memory.
(The author has long focused on AI stories at Baidu and other internet giants. Welcome to add WeChat Who123start for more insider details.)
